In finance, a confident wrong answer is the expensive kind. Same remittance, two kinds of AI:
Incoming remittance“Payment of $48,210.00 for our March invoices. Ref ACME-0321.”Open: INV-40721 $31,450 (Mar) · INV-40733 $16,510 (Mar) · INV-40790 $16,760 (Apr)
A general-purpose chatbotfluent, certain
Done. I applied the $48,210.00 payment to INV-40721 and INV-40733, the March invoices. Both are now closed.
$250 left unexplainedno confidenceno reason on file
Jev, governed by English as Codetyped, calibrated
decisionmatch_remittance
option AINV-40721 + INV-40733(Mar, $250 over)0.47
option BINV-40721 + INV-40790(exact, but Apr)0.41
your threshold0.90
“I’m not sure.” The amount fits one set of invoices and the reference fits another. Cash stays unapplied; AR review gets both options.
Illustrative example. Invoice numbers and scores are for explanation, not from a customer.
About this session
Ask most AI a question it can’t answer, and it answers anyway. In finance, that confident wrong answer is the expensive kind: cash applied to the wrong invoice, dunning paused for a customer who never promised to pay, a bank-detail change approved for a fraudster.
TypeSafe AI’s model, Jev, is built differently. It returns typed decisions with calibrated confidence, and when the evidence doesn’t support a call, it says so. Kognitos wraps Jev in English as Code, so the rules that decide what happens next, including how sure is sure enough, are written in plain English that business users, controllers, and auditors can read.
George Williams and Rushil Khurana work through what that means in practice. Roughly half the session is live: two automations built and run end to end, cash application and IT incident severity with GDPR and DORA regulatory clocks. Then they are straight about where Jev fits, where it doesn’t, and take audience questions.
Jev judges. English as Code governs.
01Jev makes the call
A typed decision (classify, match, score, route) with a calibrated confidence score.
02Your rules decide
Plain-English rules set the threshold, the action, and who reviews what. No code, no model APIs.
03Everything is on record
The input, Jev’s decision and confidence, the rule that fired, and any reviewer’s choice, in one audit trail.
You decide how sure is sure enough
Drag the threshold. Everything above it runs; everything below it goes to a person with Jev’s suggestions attached.
StepsEnglish as Code
Decide each item
If Jev is at least sure, act on its decision.
Otherwise, send the item to a reviewer with Jev’s top two suggestions.
AR email: “Invoice 40721 will be paid on the 15th”promise to pay
AR email: “We’ll sort this out next month”unclear intent
Remittance $12,400.00, one invoice, exactcash application
Remittance $48,210.00, two possible invoice setscash application
Cloud hosting invoice, known vendorGL 6100 Software
Consulting invoice, first-time vendorGL coding
Client dinner, receipt matches claimmeals & entertainment
Blurry foreign-language receiptexpense category
4 handled automatically4 sent to a person
Illustrative scores. Vendor bank-detail changes never run automatically, at any threshold.
Five finance use cases
Each one has a judgment call where a confident wrong answer is costly, so saying “I’m not sure” is the feature. In the recording, cash application is built and run live, end to end.
AR collections inbox triage
Order-to-cash
Jev decides
Classifies each customer email as a promise to pay, a dispute, a request for documents, or something else, and splits emails that cover more than one invoice.
Your rules do
Record the promise-to-pay date and pause dunning, open a dispute case for the customer’s AR owner, or send the requested documents.
“I’m not sure”
Vague or mixed emails, like “we’ll sort this out next month,” go to a reviewer with Jev’s top two guesses. It never pauses dunning on a guess.
StepsEnglish as Code
Triage each customer email
If the email is a payment commitment and Jev is at least 90% sure:
Record the promise-to-pay date.
Pause reminders for the invoice until that date.
Otherwise, send the email to the AR review queue with the top two classifications.
Cash application
Order-to-cash
Jev decides
Matches each incoming payment to open invoices from remittance text that is often partial or missing, and explains any short payment.
Your rules do
Apply cash above the threshold, book known short-pay reasons such as an early-payment discount or a freight deduction, and open a deduction case for the rest.
“I’m not sure”
When the amount and the references don’t line up with one clear set of invoices, it leaves the cash unapplied rather than forcing a match, the classic hidden error in cash application.
StepsEnglish as Code
Apply each incoming payment
If Jev matches the payment to one set of invoices and is at least 95% sure:
Apply the cash to those invoices.
Book any known short-pay reason.
Otherwise:
Leave the cash unapplied.
Open a review with the candidate matches attached.
Vendor bank-detail change requests
Procure-to-pay · fraud screen
Jev decides
Scores each request to change a vendor’s bank account for signs of payment fraud: the sender’s domain, urgency language, a mismatch with known contacts, and timing.
Your rules do
Low-risk changes go to standard callback verification. Suspicious ones freeze payments to that vendor and escalate to treasury and AP management.
“I’m not sure”
It never approves a change on its own, and an inconclusive score is treated as high risk. This is the clearest case for an AI that won’t bluff.
StepsEnglish as Code
Screen each bank-detail change
If Jev scores the request as low risk and is at least 90% sure, send it to standard callback verification.
Otherwise:
Freeze payments to the vendor.
Escalate to treasury and the AP manager.
Always
Never apply a bank-detail change without a completed callback.
AP invoice GL coding and approval routing
Procure-to-pay
Jev decides
Picks the GL account, cost center, and approver for non-PO invoices from the line descriptions and the vendor’s history.
Your rules do
Post high-confidence codings, route by amount and department, and flag invoices that break policy.
“I’m not sure”
New vendors and ambiguous line items get suggested codes for a person to confirm, not a silent miscoding that turns up at close.
StepsEnglish as Code
Code each non-PO invoice
If Jev codes the invoice and is at least 90% sure:
Post the coding.
Route it for approval by amount and department.
Otherwise, send the invoice to AP review with the top suggested codes.
Expense report policy review
Record-to-report · T&E
Jev decides
Classifies each line as a meal, travel, entertainment, or personal expense, checks receipts against claims, and flags likely policy breaches.
Your rules do
Approve compliant reports, reject clear violations with the reason, and route edge cases to the manager.
“I’m not sure”
Blurry receipts, foreign-language receipts, and unusual merchants go to review instead of being approved or rejected on a guess.
StepsEnglish as Code
Review each expense report
If every line is classified with at least 90% confidence and the report meets policy, approve the report.
If Jev finds a clear violation, reject the line with the policy reason.
Otherwise, route the report to the employee’s manager with the uncertain lines marked.
Strengths and limits, straight
What Jev is built for
Decisions under uncertainty: classifying, matching, scoring, and routing
Messy, unstructured input such as emails, remittances, and receipts
High volume where most cases are routine and a few are not
Work where you have to show an auditor why
Where it’s the wrong tool
Writing anything
Multi-step calculation
Exact rules with no judgment
Images or audio
Chapters
00:00Welcome and agenda
03:00AI is a great writer and a bad accountant
08:00How a Jev call works: Jev judges, code decides
12:00What “type-safe” actually means, and why valid is not the same as correct
33:00Live: IT incident severity with GDPR and DORA regulatory clocks
45:00The automation map
51:00Where Jev fits, and where it doesn’t
57:00Why Kognitos for Jev
59:00Audience questions
Short on time? Start at 51:00, where the session covers what Jev is the wrong tool for.
Who should watch
Controllers and Assistant Controllers
CFOs and VPs of Finance
Heads of Record-to-Report and Order-to-Cash
F&A Shared Services and GBS leaders
Internal Audit and Compliance leaders responsible for close and collections governance
Built for companies of 500 to 5,000 employees across manufacturing, logistics, healthcare, banking, retail, technology and CPG.
Want it on your own workflow?
Book a private session. Bring your messiest AR, cash application, or AP cases and we will walk through them together: what Jev would decide, where it would say it isn’t sure, and what your audit trail would look like on day one.
Common questions about this session
When was this recorded?
Wednesday, October 7, 2026, at 1:00 PM PT. The recording runs 60 minutes.
Where can I watch the recording?
The full session is embedded at the top of this page and is also published on the Kognitos YouTube channel. No registration is required.
Who are the speakers?
George Williams and Rushil Khurana.
What makes TypeSafe an honest AI?
Jev, TypeSafe AI's model, returns typed decisions with a calibrated confidence score instead of a fluent guess. When the evidence doesn't support a call, it says so, and the item goes to a person instead of being acted on.
What does English as Code add?
It makes TypeSafe usable by finance teams. The rules that decide what happens next, including how confident Jev must be before anything runs automatically, are written in plain English that controllers and auditors can read and change, with no code or model APIs.
What is built live in the session?
Two automations, end to end: cash application, and IT incident severity with GDPR and DORA regulatory clocks. Roughly half of the session is live.
What happens when the AI is not sure?
It does not act. Each decision carries a confidence score, and you set the bar it has to clear. Anything below the threshold goes to a review queue with the top suggestions attached, and the reviewer's choice is recorded alongside the original decision.
What is Jev the wrong tool for?
Writing anything, multi-step calculation, exact rules with no judgment, and images or audio. The session covers where Jev fits and where it doesn't at 51:00.
Who is this recording for?
Controllers, CFOs, AR and AP leads, and anyone evaluating decision models against a real control environment.